ship-unit
/ship-unit
Open a pull request for a completed unit, then request its external review. Verifies the unit's done-when criteria, checks that no gitignored path is staged, creates the unit branch, opens a PR whose body comes from the unit spec, and hands off to review-request. Use when a unit's work is finished and verified. Do not use mid-unit.
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 12.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Open a pull request for a completed unit, then request its external review. Verifies the unit's done-when criteria, checks that no gitignored path is staged, creates the unit branch, opens a PR whose body comes from the unit spec, and hands off to review-request. Use when a unit's work is finished and verified. Do not use mid-unit.
Original author description
Open a pull request for a completed unit, then request its external review. Verifies the unit's done-when criteria, checks that no gitignored path is staged, creates the unit branch, opens a PR whose body comes from the unit spec, and hands off to review-request. Use when a unit's work is finished and verified. Do not use mid-unit.
How it works
- 1. Identify the unit
- 2. Verify done-when
- 3. Verify nothing ignored is staged
- 4. Branch and commit
- 5. Open the PR
- 6. Report, then request the review
- Never
SKILL.md preview
--- name: ship-unit description: Open a pull request for a completed unit, then request its external review. Verifies the unit's done-when criteria, checks that no gitignored path is staged, creates the unit branch, opens a PR whose body comes from the unit spec, and hands off to review-request. Use when a unit's work is finished and verified. Do not use mid-unit. --- # Ship a unit Deterministic. No judgment calls. If a step fails, stop and report — never work around it. ## 1. Identify the unit Read `docs/plan/DECOMPOSITION.md`. Find the unit whose work is complete. If more than one is plausible, ask which — do not guess. Extract from its entry: scope, deliverables, out-of-scope, acceptance criteria. ## 2. Verify done-when Run every command in the unit's acceptance criteria, plus the standing four: ``` pnpm -r typecheck pnpm -r lint pnpm -r test git ls-files -- .plan/ # must print nothing: the directory is maintainer-local ``` The last check is scoped to the maintainer-local directory on purpose. The tracked planning documents live under `docs/plan/`, so a search for the word "plan" across tracked paths reports them and proves nothing. **Any failure stops the ship.** Report which criterion failed and stop. Do not open a PR with a failing gate and a note explaining it. **One exception, and only this one: the paid suites run in CI.** There are three: the two model drivers' (`@olympus-ai/driver-claude-code` and `@olympus-ai/driver-codex`), and the composed line's in `@olympus-ai/api`'s `test/paid/` (D-I1a-07), which CI runs under the same `run-driver` label and caches the same way. Everything below about a driver's report holds for each driver and for the line's report too. The model credentials exist only as CI's Actions secrets, so neither `pnpm --filter @olympus-ai/driver-claude-code test` nor `pnpm --filter @olympus-ai/driver-codex test` can run on a maintainer's machine, and any change under `packages/` leaves each driver's recorded report stale (D-A-CI-07). A unit's "the driver test runs once" criterion is met by the pull request's `run-driver` job, not locally. Run everything else with both driver packages excluded: ``` pnpm -r --no-bail --filter '!@olympus-ai/driver-claude-code' --filter '!@olympus-ai/driver-codex' test ``` That run passes when every failure in it is a paid suite's report and nothing else. Locally that report is stale, absent on a fresh clone, or the record of a run that refused for want of a key; none of those is evidence, and CI's is the only one that counts: - every failing test is a registry entry marked `(external: @olympus-ai/driver-claude-code …)`, `(external: @olympus-ai/driver-codex …)`, or `(external: @olympus-ai/api test/paid/…)`, or one of the two registry meta-tests that aggregate them (`registry > no entry is missing…` and `I8.registry-complete`) - every problem line those meta-tests print is either about an external entry of `@olympus-ai/driver-claude-code`, of `@olympus-ai/driver-codex`, or of `@olympus-ai/api`'s paid suite — it names that package's report and cites `I8.external-assertion-execution-reconciled` — or is `<claim>: missing` for a `driver.*` claim Any other failure, in any package, stops the ship as above. Say in the PR's acceptance section that the paid criteria are pending CI, and never mark them passed. The merge gate still holds: WORKFLOW step 7 merges only on green pull request checks, and those include the paid run and a fully green conformance registry against every fresh report. ## 3. Verify nothing ignored is staged ``` git status --porcelain git ls-files --ignored --exclude-standard -c # must be empty ``` Never `git add -A` or `git add .`. Stage explicit paths. ## 4. Branch and commit ``` git checkout -b unit/<id> # lowercase: unit/f3, unit/p1 ``` Commit message: `<id>: <unit title>`, then a body listing what was delivered. Reference decisions by their `docs/decisions.md` id rather than restating them. ## 5. Open the PR Base is `v2`. Never `main`. ``` gh pr create --base v2 --title "<id>: <unit title>" --body-file <tmp> ``` The body is generated from the unit spec, in this order: ```markdown ## Scope <the unit's scope, verbatim from DECOMPOSITION.md> ## Delivered <what was built, one line each> ## Out of scope for this unit <verbatim from the spec — this is what a reviewer should NOT flag as missing> ## Acceptance criteria <each criterion, with the command run and its result> ## Decisions <ids and one-line summaries from docs/decisions.md added by this unit> ## Registry delta <invariants moved between states; pending entries added or paid down. Omit this section if the unit touched no conformance assertions.> ## Review notes <anything the unit found awkward, ambiguous, or worth a second opinion — including places where the spec was wrong> ``` The out-of-scope section exists so a reviewer does not report absent work as a defect. It is the single most useful part of the body. ## 6. Report, then request the review Print the PR URL. Then invoke `review-request` for the same unit, without stopping to ask. The unit loop's next step is always the external review, so halting here only to be told to continue costs a turn and gains nothing. **Safe to chain**, because `review-request` produces two artifacts and stops: a bundle file on disk and the prompt that will be sent with it. It opens nothing, sends nothing, merges nothing, and contacts no reviewer. `review-request` then chains into `run-review` under the standing authorization in D-TOOLING-09, which holds only while the repository is public; that check is made there, at the point of sending, not here. **Only on success.** This step is reached only when step 2 passed and step 5 actually opened a pull request. A failed acceptance criterion stops at step 2 and never arrives here: a bundle built from a unit that does not pass its own gates sends a reviewer at a tree that is about to change. Do not merge. Do not push to `v2` directly. Do not invoke `run-review` yourself; `review-request` does, after its own checks. Do not begin the next unit. ## Never - Ship with a failing acceptance criterion, or reach step 6 with one - Modify a done-when criterion to make it pass - Touch `.github/workflows/`, `CLAUDE.md`, the conformance baseline, or a contract type file as part of shipping. Those are gate changes and belong to a unit that declares them in scope.
Versions
The database has no other skills named ship-unit.
Database snapshot: 13.11.2026 · package version 523dc924119b (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100create-pr
Create a branch, commit the work, and open a pull request for a completed user story, after all quality gates pass. Use when the user asks to open/create a PR or as the final step of the story workflow. - 86ship
Take a change from a branch to merged and cleaned up - open or update a pull request and see it through review, self-review the branch, merge, clean up after. Use when asked to open or update a PR, review your own branch, merge, or clean up after a task. - 844-open-pr
Opens or updates a pull request for the current feature. Invoked when the user wants to open a pull request, create a PR, update the draft PR, or runs /aido:4-open-pr. - 100dev-ship
Use to ship a finished branch. Triggers include "ship", "push", "open PR", "ship it". Runs ALL test suites, pushes, opens or updates the pull request, and watches CI until it passes or fails. Never merges. - 86lemonlime-open-pr
Open a pull request for the thread's branch. Use when the user asks for a pull request, a PR, or to send the work for review. - 100pr-open
Opens the pull request of a branch, as a draft or ready by the pull request policy, or edits its body. The body starts with a bottom line, then answers Why, What and How, in this order. How holds the plan diagrams and the visuals of the done tasks. Use in the ship step of the…